The pillar guide to AI agents in 2026 — what they are, how the agent loop works, raw API vs frameworks (LangChain, CrewAI), memory, multi-agent systems, MCP, cost, and career skills. With links to every deep-dive guide in the series.
Sub-agents let Claude Code run multiple specialized agents in parallel — one searching the codebase, another writing tests, another refactoring. This guide covers when to use them, the Task tool patterns, and the workflows that 3x output.
A complete tutorial on building an AI agent from scratch in Python — no LangChain, no framework. Just the Anthropic SDK, a tool-use loop, and ~60 lines of code that you fully understand and control.
LangChain has 80K stars. CrewAI has 20K. The raw Anthropic/OpenAI SDK is 60 lines. Which should you build your AI agent on? After shipping production agents in all three, here is the honest decision framework.
Build a working multi-agent system in Python — a coordinator agent delegates to specialized workers, handles failures, and synthesizes results. Complete code, real examples, no framework lock-in.
Gemma 4 is Google's most capable open-weight model — Apache 2.0, native function calling, Extended Thinking, and edge deployment down to Raspberry Pi. The complete builder's guide to running agentic AI locally in 2026.